The present invention relates to an image retrieval apparatus that retrieves an image, an image retrieval method, and a recording medium.
A plurality of images of the same photographic subject with taken from different directions or under different illumination conditions have image feature values differing from each other. Such variations in image feature value arise from the shape of the photographic subject, and therefore differ among respective photographic subjects. In other words, if approximated with the Gaussian distribution, the major axis varies depending on the photographic subject. When a partial vector space that is less likely to be affected by the variations in major axis direction, which is caused by a difference in photographic subject, is to be extracted, it is necessary to extract a partial space with a distribution that is as isotropic as possible by correcting the variations in major axis among different photographic subjects.
In the facial recognition system and method disclosed in JP 2005-512201 A, a pair of an input facial image and a reference facial image is projected from an input space to a higher dimension feature space, thereby obtaining more typical features of the facial image. A kernel-based Fisher face of the input facial image and the reference facial image is calculated. The input facial image and the reference facial image are indicated as points in the facial image space, and using the distance between the point of the inputted face and the point of each reference facial image, the resemblance of the input facial image to a specific facial image of the reference facial image is determined.
JP 2005-512201 A described above, however, uses non-linear conversion, and therefore has a problem of taking a long time to process. When obtained images are saved in a database in succession, for example, because non-linear conversion is conducted every time, it takes a long time to register the images, which limits the number of registration per unit time. Also, when non-linear conversion is conducted on the obtained images, image feature values, which are used to indicate the degree of similarity between images, cannot be obtained.